arch/x86/kernel/nmi_32.c

   1 /*
   2  *  NMI watchdog support on APIC systems
   3  *
   4  *  Started by Ingo Molnar <mingo@redhat.com>
   5  *
   6  *  Fixes:
   7  *  Mikael Pettersson   : AMD K7 support for local APIC NMI watchdog.
   8  *  Mikael Pettersson   : Power Management for local APIC NMI watchdog.
   9  *  Mikael Pettersson   : Pentium 4 support for local APIC NMI watchdog.
  10  *  Pavel Machek and
  11  *  Mikael Pettersson   : PM converted to driver model. Disable/enable API.
  12  */
  13
  14 #include <linux/delay.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/module.h>
  17 #include <linux/nmi.h>
  18 #include <linux/sysdev.h>
  19 #include <linux/sysctl.h>
  20 #include <linux/percpu.h>
  21 #include <linux/kprobes.h>
  22 #include <linux/cpumask.h>
  23 #include <linux/kernel_stat.h>
  24 #include <linux/kdebug.h>
  25
  26 #include <asm/smp.h>
  27 #include <asm/nmi.h>
  28 #include <asm/timer.h>
  29
  30 #include "mach_traps.h"
  31
  32 int unknown_nmi_panic;
  33 int nmi_watchdog_enabled;
  34
  35 static cpumask_t backtrace_mask = CPU_MASK_NONE;
  36
  37 /* nmi_active:
  38  * >0: the lapic NMI watchdog is active, but can be disabled
  39  * <0: the lapic NMI watchdog has not been set up, and cannot
  40  *     be enabled
  41  *  0: the lapic NMI watchdog is disabled, but can be enabled
  42  */
  43 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
  44
  45 unsigned int nmi_watchdog = NMI_DEFAULT;
  46 static unsigned int nmi_hz = HZ;
  47
  48 static DEFINE_PER_CPU(short, wd_enabled);
  49
  50 /* local prototypes */
  51 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu);
  52
  53 static int endflag __initdata = 0;
  54
  55 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
  56  * the CPU is idle. To make sure the NMI watchdog really ticks on all
  57  * CPUs during the test make them busy.
  58  */
  59 static __init void nmi_cpu_busy(void *data)
  60 {
  61 #ifdef CONFIG_SMP
  62         local_irq_enable_in_hardirq();
  63         /* Intentionally don't use cpu_relax here. This is
  64            to make sure that the performance counter really ticks,
  65            even if there is a simulator or similar that catches the
  66            pause instruction. On a real HT machine this is fine because
  67            all other CPUs are busy with "useless" delay loops and don't
  68            care if they get somewhat less cycles. */
  69         while (endflag == 0)
  70                 mb();
  71 #endif
  72 }
  73
  74 static int __init check_nmi_watchdog(void)
  75 {
  76         unsigned int *prev_nmi_count;
  77         int cpu;
  78
  79         if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
  80                 return 0;
  81
  82         if (!atomic_read(&nmi_active))
  83                 return 0;
  84
  85         prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
  86         if (!prev_nmi_count)
  87                 goto error;
  88
  89         printk(KERN_INFO "Testing NMI watchdog ... ");
  90
  91         if (nmi_watchdog == NMI_LOCAL_APIC)
  92                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
  93
  94         for_each_possible_cpu(cpu)
  95                 prev_nmi_count[cpu] = per_cpu(irq_stat, cpu).__nmi_count;
  96         local_irq_enable();
  97         mdelay((20*1000)/nmi_hz); // wait 20 ticks
  98
  99         for_each_possible_cpu(cpu) {
 100 #ifdef CONFIG_SMP
 101                 /* Check cpu_callin_map here because that is set
 102                    after the timer is started. */
 103                 if (!cpu_isset(cpu, cpu_callin_map))
 104                         continue;
 105 #endif
 106                 if (!per_cpu(wd_enabled, cpu))
 107                         continue;
 108                 if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
 109                         printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n",
 110                                 cpu,
 111                                 prev_nmi_count[cpu],
 112                                 nmi_count(cpu));
 113                         per_cpu(wd_enabled, cpu) = 0;
 114                         atomic_dec(&nmi_active);
 115                 }
 116         }
 117         endflag = 1;
 118         if (!atomic_read(&nmi_active)) {
 119                 kfree(prev_nmi_count);
 120                 atomic_set(&nmi_active, -1);
 121                 goto error;
 122         }
 123         printk("OK.\n");
 124
 125         /* now that we know it works we can reduce NMI frequency to
 126            something more reasonable; makes a difference in some configs */
 127         if (nmi_watchdog == NMI_LOCAL_APIC)
 128                 nmi_hz = lapic_adjust_nmi_hz(1);
 129
 130         kfree(prev_nmi_count);
 131         return 0;
 132 error:
 133         timer_ack = !cpu_has_tsc;
 134
 135         return -1;
 136 }
 137 /* This needs to happen later in boot so counters are working */
 138 late_initcall(check_nmi_watchdog);
 139
 140 static int __init setup_nmi_watchdog(char *str)
 141 {
 142         int nmi;
 143
 144         get_option(&str, &nmi);
 145
 146         if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
 147                 return 0;
 148
 149         nmi_watchdog = nmi;
 150         return 1;
 151 }
 152
 153 __setup("nmi_watchdog=", setup_nmi_watchdog);
 154
 155
 156 /* Suspend/resume support */
 157
 158 #ifdef CONFIG_PM
 159
 160 static int nmi_pm_active; /* nmi_active before suspend */
 161
 162 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
 163 {
 164         /* only CPU0 goes here, other CPUs should be offline */
 165         nmi_pm_active = atomic_read(&nmi_active);
 166         stop_apic_nmi_watchdog(NULL);
 167         BUG_ON(atomic_read(&nmi_active) != 0);
 168         return 0;
 169 }
 170
 171 static int lapic_nmi_resume(struct sys_device *dev)
 172 {
 173         /* only CPU0 goes here, other CPUs should be offline */
 174         if (nmi_pm_active > 0) {
 175                 setup_apic_nmi_watchdog(NULL);
 176                 touch_nmi_watchdog();
 177         }
 178         return 0;
 179 }
 180
 181
 182 static struct sysdev_class nmi_sysclass = {
 183         set_kset_name("lapic_nmi"),
 184         .resume         = lapic_nmi_resume,
 185         .suspend        = lapic_nmi_suspend,
 186 };
 187
 188 static struct sys_device device_lapic_nmi = {
 189         .id     = 0,
 190         .cls    = &nmi_sysclass,
 191 };
 192
 193 static int __init init_lapic_nmi_sysfs(void)
 194 {
 195         int error;
 196
 197         /* should really be a BUG_ON but b/c this is an
 198          * init call, it just doesn't work.  -dcz
 199          */
 200         if (nmi_watchdog != NMI_LOCAL_APIC)
 201                 return 0;
 202
 203         if (atomic_read(&nmi_active) < 0)
 204                 return 0;
 205
 206         error = sysdev_class_register(&nmi_sysclass);
 207         if (!error)
 208                 error = sysdev_register(&device_lapic_nmi);
 209         return error;
 210 }
 211 /* must come after the local APIC's device_initcall() */
 212 late_initcall(init_lapic_nmi_sysfs);
 213
 214 #endif  /* CONFIG_PM */
 215
 216 static void __acpi_nmi_enable(void *__unused)
 217 {
 218         apic_write_around(APIC_LVT0, APIC_DM_NMI);
 219 }
 220
 221 /*
 222  * Enable timer based NMIs on all CPUs:
 223  */
 224 void acpi_nmi_enable(void)
 225 {
 226         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 227                 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
 228 }
 229
 230 static void __acpi_nmi_disable(void *__unused)
 231 {
 232         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
 233 }
 234
 235 /*
 236  * Disable timer based NMIs on all CPUs:
 237  */
 238 void acpi_nmi_disable(void)
 239 {
 240         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
 241                 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
 242 }
 243
 244 void setup_apic_nmi_watchdog (void *unused)
 245 {
 246         if (__get_cpu_var(wd_enabled))
 247                 return;
 248
 249         /* cheap hack to support suspend/resume */
 250         /* if cpu0 is not active neither should the other cpus */
 251         if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
 252                 return;
 253
 254         switch (nmi_watchdog) {
 255         case NMI_LOCAL_APIC:
 256                 __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
 257                 if (lapic_watchdog_init(nmi_hz) < 0) {
 258                         __get_cpu_var(wd_enabled) = 0;
 259                         return;
 260                 }
 261                 /* FALL THROUGH */
 262         case NMI_IO_APIC:
 263                 __get_cpu_var(wd_enabled) = 1;
 264                 atomic_inc(&nmi_active);
 265         }
 266 }
 267
 268 void stop_apic_nmi_watchdog(void *unused)
 269 {
 270         /* only support LOCAL and IO APICs for now */
 271         if ((nmi_watchdog != NMI_LOCAL_APIC) &&
 272             (nmi_watchdog != NMI_IO_APIC))
 273                 return;
 274         if (__get_cpu_var(wd_enabled) == 0)
 275                 return;
 276         if (nmi_watchdog == NMI_LOCAL_APIC)
 277                 lapic_watchdog_stop();
 278         __get_cpu_var(wd_enabled) = 0;
 279         atomic_dec(&nmi_active);
 280 }
 281
 282 /*
 283  * the best way to detect whether a CPU has a 'hard lockup' problem
 284  * is to check it's local APIC timer IRQ counts. If they are not
 285  * changing then that CPU has some problem.
 286  *
 287  * as these watchdog NMI IRQs are generated on every CPU, we only
 288  * have to check the current processor.
 289  *
 290  * since NMIs don't listen to _any_ locks, we have to be extremely
 291  * careful not to rely on unsafe variables. The printk might lock
 292  * up though, so we have to break up any console locks first ...
 293  * [when there will be more tty-related locks, break them up
 294  *  here too!]
 295  */
 296
 297 static unsigned int
 298         last_irq_sums [NR_CPUS],
 299         alert_counter [NR_CPUS];
 300
 301 void touch_nmi_watchdog(void)
 302 {
 303         if (nmi_watchdog > 0) {
 304                 unsigned cpu;
 305
 306                 /*
 307                  * Just reset the alert counters, (other CPUs might be
 308                  * spinning on locks we hold):
 309                  */
 310                 for_each_present_cpu(cpu) {
 311                         if (alert_counter[cpu])
 312                                 alert_counter[cpu] = 0;
 313                 }
 314         }
 315
 316         /*
 317          * Tickle the softlockup detector too:
 318          */
 319         touch_softlockup_watchdog();
 320 }
 321 EXPORT_SYMBOL(touch_nmi_watchdog);
 322
 323 extern void die_nmi(struct pt_regs *, const char *msg);
 324
 325 __kprobes int nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
 326 {
 327
 328         /*
 329          * Since current_thread_info()-> is always on the stack, and we
 330          * always switch the stack NMI-atomically, it's safe to use
 331          * smp_processor_id().
 332          */
 333         unsigned int sum;
 334         int touched = 0;
 335         int cpu = smp_processor_id();
 336         int rc=0;
 337
 338         /* check for other users first */
 339         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
 340                         == NOTIFY_STOP) {
 341                 rc = 1;
 342                 touched = 1;
 343         }
 344
 345         if (cpu_isset(cpu, backtrace_mask)) {
 346                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
 347
 348                 spin_lock(&lock);
 349                 printk("NMI backtrace for cpu %d\n", cpu);
 350                 dump_stack();
 351                 spin_unlock(&lock);
 352                 cpu_clear(cpu, backtrace_mask);
 353         }
 354
 355         /*
 356          * Take the local apic timer and PIT/HPET into account. We don't
 357          * know which one is active, when we have highres/dyntick on
 358          */
 359         sum = per_cpu(irq_stat, cpu).apic_timer_irqs +
 360                 per_cpu(irq_stat, cpu).irq0_irqs;
 361
 362         /* if the none of the timers isn't firing, this cpu isn't doing much */
 363         if (!touched && last_irq_sums[cpu] == sum) {
 364                 /*
 365                  * Ayiee, looks like this CPU is stuck ...
 366                  * wait a few IRQs (5 seconds) before doing the oops ...
 367                  */
 368                 alert_counter[cpu]++;
 369                 if (alert_counter[cpu] == 5*nmi_hz)
 370                         /*
 371                          * die_nmi will return ONLY if NOTIFY_STOP happens..
 372                          */
 373                         die_nmi(regs, "BUG: NMI Watchdog detected LOCKUP");
 374         } else {
 375                 last_irq_sums[cpu] = sum;
 376                 alert_counter[cpu] = 0;
 377         }
 378         /* see if the nmi watchdog went off */
 379         if (!__get_cpu_var(wd_enabled))
 380                 return rc;
 381         switch (nmi_watchdog) {
 382         case NMI_LOCAL_APIC:
 383                 rc |= lapic_wd_event(nmi_hz);
 384                 break;
 385         case NMI_IO_APIC:
 386                 /* don't know how to accurately check for this.
 387                  * just assume it was a watchdog timer interrupt
 388                  * This matches the old behaviour.
 389                  */
 390                 rc = 1;
 391                 break;
 392         }
 393         return rc;
 394 }
 395
 396 int do_nmi_callback(struct pt_regs * regs, int cpu)
 397 {
 398 #ifdef CONFIG_SYSCTL
 399         if (unknown_nmi_panic)
 400                 return unknown_nmi_panic_callback(regs, cpu);
 401 #endif
 402         return 0;
 403 }
 404
 405 #ifdef CONFIG_SYSCTL
 406
 407 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
 408 {
 409         unsigned char reason = get_nmi_reason();
 410         char buf[64];
 411
 412         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
 413         die_nmi(regs, buf);
 414         return 0;
 415 }
 416
 417 /*
 418  * proc handler for /proc/sys/kernel/nmi
 419  */
 420 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
 421                         void __user *buffer, size_t *length, loff_t *ppos)
 422 {
 423         int old_state;
 424
 425         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
 426         old_state = nmi_watchdog_enabled;
 427         proc_dointvec(table, write, file, buffer, length, ppos);
 428         if (!!old_state == !!nmi_watchdog_enabled)
 429                 return 0;
 430
 431         if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
 432                 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
 433                 return -EIO;
 434         }
 435
 436         if (nmi_watchdog == NMI_DEFAULT) {
 437                 if (lapic_watchdog_ok())
 438                         nmi_watchdog = NMI_LOCAL_APIC;
 439                 else
 440                         nmi_watchdog = NMI_IO_APIC;
 441         }
 442
 443         if (nmi_watchdog == NMI_LOCAL_APIC) {
 444                 if (nmi_watchdog_enabled)
 445                         enable_lapic_nmi_watchdog();
 446                 else
 447                         disable_lapic_nmi_watchdog();
 448         } else {
 449                 printk( KERN_WARNING
 450                         "NMI watchdog doesn't know what hardware to touch\n");
 451                 return -EIO;
 452         }
 453         return 0;
 454 }
 455
 456 #endif
 457
 458 void __trigger_all_cpu_backtrace(void)
 459 {
 460         int i;
 461
 462         backtrace_mask = cpu_online_map;
 463         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
 464         for (i = 0; i < 10 * 1000; i++) {
 465                 if (cpus_empty(backtrace_mask))
 466                         break;
 467                 mdelay(1);
 468         }
 469 }
 470
 471 EXPORT_SYMBOL(nmi_active);
 472 EXPORT_SYMBOL(nmi_watchdog);